nmi.c 6.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286
  1. /* Pseudo NMI support on sparc64 systems.
  2. *
  3. * Copyright (C) 2009 David S. Miller <davem@davemloft.net>
  4. *
  5. * The NMI watchdog support and infrastructure is based almost
  6. * entirely upon the x86 NMI support code.
  7. */
  8. #include <linux/kernel.h>
  9. #include <linux/param.h>
  10. #include <linux/init.h>
  11. #include <linux/percpu.h>
  12. #include <linux/nmi.h>
  13. #include <linux/export.h>
  14. #include <linux/kprobes.h>
  15. #include <linux/kernel_stat.h>
  16. #include <linux/reboot.h>
  17. #include <linux/slab.h>
  18. #include <linux/kdebug.h>
  19. #include <linux/delay.h>
  20. #include <linux/smp.h>
  21. #include <asm/perf_event.h>
  22. #include <asm/ptrace.h>
  23. #include <asm/pcr.h>
  24. #include <asm/perfctr.h>
  25. #include "kstack.h"
  26. /* We don't have a real NMI on sparc64, but we can fake one
  27. * up using profiling counter overflow interrupts and interrupt
  28. * levels.
  29. *
  30. * The profile overflow interrupts at level 15, so we use
  31. * level 14 as our IRQ off level.
  32. */
  33. static int panic_on_timeout;
  34. /* nmi_active:
  35. * >0: the NMI watchdog is active, but can be disabled
  36. * <0: the NMI watchdog has not been set up, and cannot be enabled
  37. * 0: the NMI watchdog is disabled, but can be enabled
  38. */
  39. atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */
  40. EXPORT_SYMBOL(nmi_active);
  41. static unsigned int nmi_hz = HZ;
  42. static DEFINE_PER_CPU(short, wd_enabled);
  43. static int endflag __initdata;
  44. static DEFINE_PER_CPU(unsigned int, last_irq_sum);
  45. static DEFINE_PER_CPU(long, alert_counter);
  46. static DEFINE_PER_CPU(int, nmi_touch);
  47. void touch_nmi_watchdog(void)
  48. {
  49. if (atomic_read(&nmi_active)) {
  50. int cpu;
  51. for_each_present_cpu(cpu) {
  52. if (per_cpu(nmi_touch, cpu) != 1)
  53. per_cpu(nmi_touch, cpu) = 1;
  54. }
  55. }
  56. touch_softlockup_watchdog();
  57. }
  58. EXPORT_SYMBOL(touch_nmi_watchdog);
  59. static void die_nmi(const char *str, struct pt_regs *regs, int do_panic)
  60. {
  61. if (notify_die(DIE_NMIWATCHDOG, str, regs, 0,
  62. pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
  63. return;
  64. console_verbose();
  65. bust_spinlocks(1);
  66. printk(KERN_EMERG "%s", str);
  67. printk(" on CPU%d, ip %08lx, registers:\n",
  68. smp_processor_id(), regs->tpc);
  69. show_regs(regs);
  70. dump_stack();
  71. bust_spinlocks(0);
  72. if (do_panic || panic_on_oops)
  73. panic("Non maskable interrupt");
  74. nmi_exit();
  75. local_irq_enable();
  76. do_exit(SIGBUS);
  77. }
  78. notrace __kprobes void perfctr_irq(int irq, struct pt_regs *regs)
  79. {
  80. unsigned int sum, touched = 0;
  81. void *orig_sp;
  82. clear_softint(1 << irq);
  83. local_cpu_data().__nmi_count++;
  84. nmi_enter();
  85. orig_sp = set_hardirq_stack();
  86. if (notify_die(DIE_NMI, "nmi", regs, 0,
  87. pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
  88. touched = 1;
  89. else
  90. pcr_ops->write(PCR_PIC_PRIV);
  91. sum = local_cpu_data().irq0_irqs;
  92. if (__get_cpu_var(nmi_touch)) {
  93. __get_cpu_var(nmi_touch) = 0;
  94. touched = 1;
  95. }
  96. if (!touched && __get_cpu_var(last_irq_sum) == sum) {
  97. __this_cpu_inc(alert_counter);
  98. if (__this_cpu_read(alert_counter) == 30 * nmi_hz)
  99. die_nmi("BUG: NMI Watchdog detected LOCKUP",
  100. regs, panic_on_timeout);
  101. } else {
  102. __get_cpu_var(last_irq_sum) = sum;
  103. __this_cpu_write(alert_counter, 0);
  104. }
  105. if (__get_cpu_var(wd_enabled)) {
  106. write_pic(picl_value(nmi_hz));
  107. pcr_ops->write(pcr_enable);
  108. }
  109. restore_hardirq_stack(orig_sp);
  110. nmi_exit();
  111. }
  112. static inline unsigned int get_nmi_count(int cpu)
  113. {
  114. return cpu_data(cpu).__nmi_count;
  115. }
  116. static __init void nmi_cpu_busy(void *data)
  117. {
  118. local_irq_enable_in_hardirq();
  119. while (endflag == 0)
  120. mb();
  121. }
  122. static void report_broken_nmi(int cpu, int *prev_nmi_count)
  123. {
  124. printk(KERN_CONT "\n");
  125. printk(KERN_WARNING
  126. "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
  127. cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
  128. printk(KERN_WARNING
  129. "Please report this to bugzilla.kernel.org,\n");
  130. printk(KERN_WARNING
  131. "and attach the output of the 'dmesg' command.\n");
  132. per_cpu(wd_enabled, cpu) = 0;
  133. atomic_dec(&nmi_active);
  134. }
  135. void stop_nmi_watchdog(void *unused)
  136. {
  137. pcr_ops->write(PCR_PIC_PRIV);
  138. __get_cpu_var(wd_enabled) = 0;
  139. atomic_dec(&nmi_active);
  140. }
  141. static int __init check_nmi_watchdog(void)
  142. {
  143. unsigned int *prev_nmi_count;
  144. int cpu, err;
  145. if (!atomic_read(&nmi_active))
  146. return 0;
  147. prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(unsigned int), GFP_KERNEL);
  148. if (!prev_nmi_count) {
  149. err = -ENOMEM;
  150. goto error;
  151. }
  152. printk(KERN_INFO "Testing NMI watchdog ... ");
  153. smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
  154. for_each_possible_cpu(cpu)
  155. prev_nmi_count[cpu] = get_nmi_count(cpu);
  156. local_irq_enable();
  157. mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
  158. for_each_online_cpu(cpu) {
  159. if (!per_cpu(wd_enabled, cpu))
  160. continue;
  161. if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
  162. report_broken_nmi(cpu, prev_nmi_count);
  163. }
  164. endflag = 1;
  165. if (!atomic_read(&nmi_active)) {
  166. kfree(prev_nmi_count);
  167. atomic_set(&nmi_active, -1);
  168. err = -ENODEV;
  169. goto error;
  170. }
  171. printk("OK.\n");
  172. nmi_hz = 1;
  173. kfree(prev_nmi_count);
  174. return 0;
  175. error:
  176. on_each_cpu(stop_nmi_watchdog, NULL, 1);
  177. return err;
  178. }
  179. void start_nmi_watchdog(void *unused)
  180. {
  181. __get_cpu_var(wd_enabled) = 1;
  182. atomic_inc(&nmi_active);
  183. pcr_ops->write(PCR_PIC_PRIV);
  184. write_pic(picl_value(nmi_hz));
  185. pcr_ops->write(pcr_enable);
  186. }
  187. static void nmi_adjust_hz_one(void *unused)
  188. {
  189. if (!__get_cpu_var(wd_enabled))
  190. return;
  191. pcr_ops->write(PCR_PIC_PRIV);
  192. write_pic(picl_value(nmi_hz));
  193. pcr_ops->write(pcr_enable);
  194. }
  195. void nmi_adjust_hz(unsigned int new_hz)
  196. {
  197. nmi_hz = new_hz;
  198. on_each_cpu(nmi_adjust_hz_one, NULL, 1);
  199. }
  200. EXPORT_SYMBOL_GPL(nmi_adjust_hz);
  201. static int nmi_shutdown(struct notifier_block *nb, unsigned long cmd, void *p)
  202. {
  203. on_each_cpu(stop_nmi_watchdog, NULL, 1);
  204. return 0;
  205. }
  206. static struct notifier_block nmi_reboot_notifier = {
  207. .notifier_call = nmi_shutdown,
  208. };
  209. int __init nmi_init(void)
  210. {
  211. int err;
  212. on_each_cpu(start_nmi_watchdog, NULL, 1);
  213. err = check_nmi_watchdog();
  214. if (!err) {
  215. err = register_reboot_notifier(&nmi_reboot_notifier);
  216. if (err) {
  217. on_each_cpu(stop_nmi_watchdog, NULL, 1);
  218. atomic_set(&nmi_active, -1);
  219. }
  220. }
  221. return err;
  222. }
  223. static int __init setup_nmi_watchdog(char *str)
  224. {
  225. if (!strncmp(str, "panic", 5))
  226. panic_on_timeout = 1;
  227. return 0;
  228. }
  229. __setup("nmi_watchdog=", setup_nmi_watchdog);